# Claude Opus 5 lowers the cost of frontier agent work, but coding gains are uneven > Source: > Published: 2026-07-26 13:29:57+00:00 # Claude Opus 5 lowers the cost of frontier agent work, but coding gains are uneven - Opus 5 costs $5 per million input tokens and $25 per million output tokens, half Fable 5’s list rates and unchanged from Opus 4.8. [[1]](https://www.anthropic.com/news/claude-opus-5)[[2]](https://platform.claude.com/docs/en/about-claude/pricing) - Artificial Analysis placed Opus 5 narrowly atop its Intelligence Index, but maximum-effort runs still cost more than Opus 4.8 and Sonnet 5 on that test. [[3]](https://artificialanalysis.ai/articles/opus-5) - CodeRabbit found higher precision but lower issue coverage and more low-value comments in one Opus 5 code-review configuration. [[4]](https://www.coderabbit.ai/blog/opus-5-model-review) - Opus 5 is available through Anthropic, AWS and Google Cloud, with a gradual rollout to paid GitHub Copilot plans. [[5]](https://aws.amazon.com/blogs/machine-learning/introducing-claude-opus-5-on-aws-anthropics-most-capable-opus-model/)[[6]](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5)[[7]](https://github.blog/changelog/2026-07-24-claude-opus-5-is-now-available-in-github-copilot/) Anthropic released Claude Opus 5 on July 24, pitching it as a lower-cost alternative to its Fable 5 frontier model. The model costs $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8 and half Fable 5’s $10 and $50 list rates.[[1]](https://www.anthropic.com/news/claude-opus-5)[[2]](https://platform.claude.com/docs/en/about-claude/pricing) The release is primarily an efficiency upgrade. Anthropic says Opus 5 more than doubles Opus 4.8’s performance on its Frontier-Bench v0.1 run and comes within 0.5% of Fable 5’s peak CursorBench 3.2 score at half the cost per task. Anthropic’s Frontier-Bench setup allowed Opus 4.8 to handle requests refused by Opus 5’s safety classifier, however, meaning the deployed system—not only the new model—contributed to that result.[[1]](https://www.anthropic.com/news/claude-opus-5) Independent results broadly support the cost-performance pitch without showing a clean win everywhere. Artificial Analysis ranked maximum-effort Opus 5 narrowly ahead of Fable 5 on its composite Intelligence Index, while CodeRabbit found that one high-effort code-review setup traded better precision for lower coverage and more noise.[[3]](https://artificialanalysis.ai/articles/opus-5)[[4]](https://www.coderabbit.ai/blog/opus-5-model-review) ## The savings depend on effort and workload Artificial Analysis reported that maximum-effort Opus 5 scored 61 on its Intelligence Index, effectively tied with Fable 5 at 60 and ahead of Opus 4.8 at 56. The test cost an estimated $2.03 per task for Opus 5, below Fable 5’s $2.75 but above Opus 4.8’s $1.80 and Sonnet 5’s $1.53. Artificial Analysis ran its Opus 5 evaluations with Opus 4.8 fallback enabled, another reason to interpret the numbers as measurements of a configured agent system rather than an isolated model.[[3]](https://artificialanalysis.ai/articles/opus-5) On AA-Briefcase, Artificial Analysis’ private benchmark for multi-file knowledge work, maximum-effort Opus 5 scored 1,720 Elo, 146 points above Fable 5. Its estimated cost was $17.79 per task, 20% below Fable’s $22.30. At high effort, Opus 5 remained 32 Elo ahead of Fable while costing $10.41 per task.[[8]](https://artificialanalysis.ai/articles/claude-opus-5-leader-agentic-knowledge-work) Higher scores required time. AA-Briefcase tasks averaged 36.2 minutes at maximum effort, 34.3 minutes at extra-high effort and 25.7 minutes at high effort. Maximum-effort runs averaged 103 turns, compared with 55 for Opus 4.8. The model’s effort control is therefore a routing decision, not a simple quality switch: more reasoning can raise performance while increasing latency and token consumption.[[8]](https://artificialanalysis.ai/articles/claude-opus-5-leader-agentic-knowledge-work) Anthropic also offers a Fast mode that it says runs about 2.5 times as quickly at twice the base API price. Sonnet 5 is the cheaper alternative at $2 per million input tokens and $10 per million output tokens through August 31, rising to $3 and $15 on September 1.[[1]](https://www.anthropic.com/news/claude-opus-5)[[2]](https://platform.claude.com/docs/en/about-claude/pricing) ## Coding-agent performance is not a uniform upgrade Anthropic emphasizes long-running software work, including debugging, root-cause analysis and verification. Artificial Analysis placed Opus 5 with Claude Code in a joint-leading position on its Coding Agent Index and measured an 89% maximum-effort score on Terminal-Bench 2.1, roughly in line with GPT-5.6 Sol at extra-high effort.[[3]](https://artificialanalysis.ai/articles/opus-5) CodeRabbit reached a narrower conclusion in a test covering about 100 error patterns taken from verified issues in open-source pull requests. Its extra-high-effort Opus 5 configuration produced actionable comments with 39.3% precision, above its 35.2% production baseline. It caught 55.2% of known issues, however, compared with 61.1% for the baseline, and produced 92 nitpicks versus 23.[[4]](https://www.coderabbit.ai/blog/opus-5-model-review) That result does not establish how Opus 5 will perform across coding agents or full software-development cycles. It does show why teams should evaluate precision, coverage, review burden and cost separately. CodeRabbit concluded that the tested configuration was better suited to serving as a second, precision-oriented reviewer than as the only safety net for high-risk changes.[[4]](https://www.coderabbit.ai/blog/opus-5-model-review) Artificial Analysis also found that Opus 5’s factual knowledge remained below Fable 5 on its AA-Omniscience evaluation. Although accuracy improved by seven points over Opus 4.8, the measured hallucination rate rose by 14 points to 50%. The result reinforces that better agentic performance does not imply an across-the-board improvement in factual reliability.[[3]](https://artificialanalysis.ai/articles/opus-5) ## Enterprise access comes with safety routing Opus 5 is the default model for Claude Max and the strongest model offered through Claude Pro. AWS made it available through Amazon Bedrock and Claude Platform on AWS, including regional deployments and zero-data-retention controls. Google Cloud lists the model as generally available with a one-million-token input limit, a 128,000-token output limit and US, European and global endpoints.[[1]](https://www.anthropic.com/news/claude-opus-5)[[5]](https://aws.amazon.com/blogs/machine-learning/introducing-claude-opus-5-on-aws-anthropics-most-capable-opus-model/)[[6]](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5) GitHub is gradually rolling out Opus 5 to Copilot Pro+, Max, Business and Enterprise users across supported editors, its command-line interface, mobile apps and the Copilot cloud agent. Business and Enterprise administrators must enable the model in Copilot settings.[[7]](https://github.blog/changelog/2026-07-24-claude-opus-5-is-now-available-in-github-copilot/) Anthropic says Opus 5 does not advance the frontier in offensive cybersecurity or biological capabilities. Its classifiers permit source-code vulnerability discovery but block binary vulnerability scanning, penetration testing and exploit generation. The company estimates that these classifiers will intervene about 85% less often than Fable 5’s safeguards.[[1]](https://www.anthropic.com/news/claude-opus-5) Flagged requests in Claude, Claude Code and Cowork fall back to Opus 4.8 by default; API customers can enable the same behavior. GitHub separately warns that some security-adjacent prompts may be blocked. Enterprises evaluating security-related coding work should log fallback events because the model selected by the user may not be the model that completes the task.[[1]](https://www.anthropic.com/news/claude-opus-5)[[7]](https://github.blog/changelog/2026-07-24-claude-opus-5-is-now-available-in-github-copilot/) ## Companies mentioned ## Further sources [[1] Anthropic, “Introducing Claude Opus 5,” July 24, 2026. ↗](https://www.anthropic.com/news/claude-opus-5) [[2] Anthropic Claude Platform documentation, “Pricing,” accessed July 26, 2026. ↗](https://platform.claude.com/docs/en/about-claude/pricing) [[3] Artificial Analysis, “Opus 5: Fable 5 level intelligence at a lower cost per ta… ↗](https://artificialanalysis.ai/articles/opus-5) [[4] CodeRabbit, “Opus 5 for code review: Cleaner actionable comments, noisier overa… ↗](https://www.coderabbit.ai/blog/opus-5-model-review) [[5] Amazon Web Services, “Introducing Claude Opus 5 on AWS,” July 24, 2026. ↗](https://aws.amazon.com/blogs/machine-learning/introducing-claude-opus-5-on-aws-anthropics-most-capable-opus-model/) [[6] Google Cloud documentation, “Claude Opus 5 on Google Cloud,” updated July 24, 2… ↗](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/opus-5)+2 more The stories that matter, in one email. Free — unsubscribe anytime.